On-Device AI Feature Budget: What US Enterprise Teams Spend to Ship Local Inference in 2026

On-device AI costs more to build than cloud AI. That is true and it matters. But most enterprise teams are quoted a number without understanding what it buys. This is what text AI, voice AI, and image AI each cost to ship on-device — broken down by phase, with the variables that shift estimates up or down.

What drives on-device AI build cost

Five factors determine where a given on-device AI project lands in the cost range.

Feature complexity. A text AI feature that classifies short documents is simpler than one that generates multi-paragraph responses in context of a long conversation. More complex features require larger models, more prompt engineering, and more edge case testing.

Platform scope. An iOS-only app at a premium device floor (A15+ chip, 2022+) is the cheapest scenario. A cross-platform app targeting Android and iOS, including mid-range devices from 2020 forward, is the most expensive. Wednesday ships Off Grid on iOS, Android, and macOS from a single React Native app — that scope requires more backend complexity than a single platform.

Device compatibility floor. Setting the minimum supported device to 2023+ flagship hardware allows smaller models and simpler inference backends. Supporting 2019+ devices requires more optimization work, lower model sizes, and additional fallback logic. Most enterprise apps targeting current users can set a 2021+ floor, which reduces complexity meaningfully.

Model selection. Choosing a model that fits comfortably in a device's RAM with acceptable latency requires benchmarking across the target device matrix. This is not optional work — a model that performs well on a test device may perform poorly on a mid-range device in the field.

QA scope. On-device AI produces different outputs than cloud AI for the same inputs, and those outputs vary by device. QA for on-device AI requires testing the AI feature specifically on each hardware platform, not just functional testing of the interface.

Text AI budget breakdown

Text AI covers features that generate, classify, summarise, or extract text locally on the device. Examples: documentation assistants, smart search, note summarisation, form extraction, and classification of support tickets or work orders.

The inference engine for text AI is llama.cpp — a highly optimised C++ library that runs quantised large language models on CPU. For devices with available hardware acceleration, Core ML (iOS) and QNN (Snapdragon Android) provide faster inference.

Phase-by-phase costs:

Model selection and benchmarking: $15,000-$30,000 This is often underestimated. Choosing the right model means defining the benchmark tasks specific to your use case, running candidates (Llama 3 8B, Phi-4, Mistral 7B, Gemma 2 9B) through those tasks, evaluating output quality, measuring inference speed and RAM usage on your target device matrix, and selecting the model that balances quality, speed, and device compatibility. For enterprise apps with specialised vocabulary or strict accuracy requirements, fine-tuning an open-source base model adds cost.

Integration engineering: $25,000-$60,000 Building the inference bridge into the mobile app. For React Native apps, this is a C++ native module that exposes the llama.cpp interface to JavaScript. For native iOS, it is a Swift wrapper. For native Android, it is a Kotlin/JNI bridge. Includes model download and caching logic (the model weights are typically 2-6GB and cannot ship in the app binary), streaming output for responsive UI, and context management for multi-turn conversations.

Device compatibility QA: $10,000-$20,000 Testing the feature on a physical device matrix — not just simulators. Minimum: flagship iOS (A15+), mid-range iOS (A13), flagship Android (Snapdragon 8 Gen 2), mid-range Android (Snapdragon 778G or equivalent). Each device may exhibit different inference speed, occasional generation artifacts, or unexpected memory pressure. Edge cases discovered here require engineering fixes.

Ongoing model updates: $5,000-$15,000 per year Annual model updates are optional but recommended. Each update cycle involves re-benchmarking, updating the bundled model reference, regression testing the AI feature, and shipping an app update.

Total text AI budget: $55,000-$130,000 build, plus $5,000-$15,000 annually.

Voice transcription budget breakdown

Voice AI covers on-device transcription using Whisper — the open-source speech recognition model that runs locally without sending audio to a server.

Model selection: $5,000-$10,000 Whisper has multiple size variants: tiny (39MB), base (74MB), small (244MB), medium (769MB), large (1.5GB). Selecting the right model for your accuracy requirements and device floor requires listening tests on representative audio samples from your actual use case — field recordings, clinical dictation, or phone audio.

Integration engineering: $20,000-$40,000 Whisper requires an audio capture pipeline, chunking logic (to feed audio to the model in appropriate segments), post-processing to clean up transcription output, and UI to display partial results as transcription progresses. For React Native apps, this is a native module. For native apps, it is a direct integration.

QA: $5,000-$10,000 Testing accuracy across accents, noise levels, and recording conditions representative of your users' environments. Field service apps need testing in noisy industrial environments. Clinical apps need testing with medical terminology. Sales apps need testing with phone audio quality.

Total voice AI budget: $30,000-$60,000 build, plus $3,000-$8,000 annually.

Image AI budget breakdown

On-device image generation is the most complex and expensive category. The complexity comes from the number of hardware backends required to cover the device matrix.

ARM64 Android devices without Snapdragon NPU hardware acceleration run image generation through MNN — a mobile neural network inference framework from Alibaba. Snapdragon 8 Gen 1+ devices can use Qualcomm's QNN framework to access the NPU, producing faster inference with lower battery impact. iOS devices use Core ML, Apple's native inference framework with Metal GPU acceleration.

A production on-device image generation feature requires all three backends: MNN for the Android long tail, QNN for flagship Android, and Core ML for iOS. Each backend produces slightly different outputs at the edges. Each requires independent QA.

Model selection and adaptation: $15,000-$30,000 Selecting a diffusion model (SDXL Turbo, LCM, or similar) that fits within the RAM constraints of the target device floor, benchmarking generation speed and quality across backends, and adapting model weights for each inference framework.

Multi-backend integration engineering: $50,000-$100,000 Building the three inference backends, unifying them behind a single JavaScript API (for React Native) or platform abstraction layer (for native), handling backend selection logic at runtime, and building the image output pipeline including progressive display and post-processing.

Cross-platform QA: $15,000-$30,000 Testing image generation quality, speed, and error handling across the full device matrix. Image generation edge cases are hardware-specific and require physical device testing.

Total image AI budget: $80,000-$160,000 build, plus $10,000-$20,000 annually.

Device compatibility: the overlooked cost

Device compatibility is the single most variable cost driver in on-device AI builds. Enterprise teams that scope on the assumption that every user has a 2023 flagship device are consistently surprised by reality.

Enterprise device fleets are heterogeneous. A field service company with 2,000 technicians may have devices ranging from 2019 Samsung A-series to 2024 iPhone 15. A clinical organisation may have a standardised device fleet but on a 5-year replacement cycle. A financial services firm may allow BYOD, meaning the device matrix is unbounded.

For each older device that needs to be supported, the engineering choices are:

• Smaller models that fit in less RAM (reducing quality)
• CPU-only inference without NPU acceleration (increasing latency)
• Graceful degradation that shows a simplified feature on older hardware
• A minimum device specification that excludes older devices from the AI feature

Each choice adds scope to the specification. Wednesday recommends defining the device floor in the first week of any on-device AI project — before any engineering begins — and communicating it to the enterprise client so device upgrade planning can happen alongside the AI feature build if needed.

Budget summary by feature type

Feature type	Build budget	Annual maintenance	On-device inference cost per query	Cloud alternative monthly at 100K DAU
Text AI (documentation, classification, Q&A)	$55,000-$130,000	$5,000-$15,000	$0	$90,000-$270,000
Voice transcription (Whisper)	$30,000-$60,000	$3,000-$8,000	$0	$240,000-$960,000
Image generation (multi-backend)	$80,000-$160,000	$10,000-$20,000	$0	$120,000-$360,000

What Off Grid proves about build cost

Wednesday built Off Grid — a complete on-device AI suite with text, voice, and image AI — as an internal product. It ships on iOS, Android, and macOS from a single React Native app. 50,000+ users. 1,700+ GitHub stars. Zero paid marketing.

The cost estimates in this article are derived from the actual engineering effort that went into Off Grid. They are not estimates assembled from general market research. They reflect the actual phases, actual complexity, and actual device compatibility work that on-device AI requires.

The implication for enterprise clients: when Wednesday scopes on-device AI work, the estimate is based on delivered work, not speculation. The integration patterns exist. The device compatibility matrix is known. The model selection process has been run. Enterprise teams are not paying for Wednesday to figure out on-device AI. They are buying a team that has already done it.

How Wednesday scopes on-device AI work

Every on-device AI scoping engagement starts with three inputs: the specific AI feature to build, the target device floor, and the data sensitivity classification.

From those inputs, Wednesday selects the appropriate model, determines which inference backends are needed, and estimates the integration and QA scope. The estimate is broken into phases — model selection, integration, QA, and maintenance — with clear deliverables at each phase.

For enterprise teams with an existing mobile app, the scoping estimate is typically produced in the first conversation. The variables that determine cost are knowable within 30 minutes of understanding the app architecture, the feature requirements, and the device matrix.